On November 19 Cognition posted Gemini 4 Neon's FrontierCode 2.0 result on X: 46.2% at $4.38 per task, the best score for any Gemini model on the benchmark. The official leaderboard puts it third, behind Claude Fable 5.2 (50.2%) and Claude Opus 5.5 (47.3%) and ahead of GPT-6.2 Astra (44.2%). Cognition says it costs about a third as much as Opus 5.5, and the model is now available in Devin Desktop and Devin CLI.
Key Takeaways
- ✓Source: an @cognition post on November 19 and the official cognition.com/frontiercode leaderboard (changelog: "Added Gemini 4 Neon")
- ✓Gemini 4 Neon: 46.2% at $4.38 per task, the highest of any Gemini model
- ✓Top of the leaderboard: Fable 5.2 50.2%, Opus 5.5 47.3%, Neon 46.2%, GPT-6.2 Astra 44.2%, GPT-6.2 Sol 43.8%, Gemini 4 Carbon 42.6%
- ✓Cognition says Neon is within about one point of Opus 5.5 at roughly a third of the cost; the post gives no exact per-task cost for Opus 5.5
- ✓Gemini 4 Neon can now be selected in Devin Desktop and Devin CLI

Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background
Google released Gemini 4 Neon on November 18, positioned between Hydrogen and Carbon, along with a set of third-party benchmark results. For teams writing code, the more practical question is whether the model's pull requests would actually be merged. FrontierCode, built by Cognition with open-source maintainers, asks exactly that across 132 tasks in version 2.0, and it is far from saturated.
How it works
A FrontierCode score is a weighted sum of rubric items, and any solution that fails a blocking check scores 0. Cost is the average spend per run at public API list prices. Models may read docs and search errors, but consulting a source that contains the answer, such as the original PR, scores 0. Cognition did not say which reasoning effort it used for Neon; the leaderboard lets you expand score and cost by effort level.
Benchmarks
According to Cognition's post, Gemini 4 Neon scores 46.2% at $4.38 per task, within about one point of Opus 5.5 at roughly a third of the cost. The official leaderboard reads: Claude Fable 5.2 50.2%, Claude Opus 5.5 47.3%, Gemini 4 Neon 46.2%, GPT-6.2 Astra 44.2%, GPT-6 Astra 44.0%, GPT-6.2 Sol 43.8%, Gemini 4 Carbon 42.6% and Claude Sonnet 5.5 42.2%. That puts Neon ahead of every current OpenAI model on this test and about 3.6 points above Carbon. These are Cognition's own results, not Google's.
What developers should do
Devin users can switch to Gemini 4 Neon in Devin Desktop and Devin CLI now. If you are picking a base model for a coding agent, Neon offers near-top scores at a clearly lower cost and is worth comparing against Opus 5.5 and Fable 5.2 on your own repositories. Still, 46.2% means most solutions fall short of the merge bar, so human review and tests remain essential.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.